关键词: 小目标检测, 医学图像, 结核杆菌, CNN

Abstract: Through sputum-smear staining, mycobacterium tuberculosis can be shown on microscope image, which makes it possible to detect M. tuberculosis on the image for facilitating tuberculosis diagnosis. On the microscope image, M. tuberculosis is characterized with diverse color saturation, various shape, and undistinguishable appearance confused with background, which make it a great challenge for traditional object detection methods. As convolutional neural networks (CNN) has achieved great success in object detection recently, we study CNN-based method, for instance, Faster-RCNN for M. tuberculosis detection. Nevertheless, there are still some problems with CNN-based tuberculosis detection:  a) Size of M. tuberculosis on image is too small, b) Constructing enough accurate labeled data is difficult, and c) Transfer learning does not work for tuberculosis detection. All of those make it hard to apply CNN-based method to M. tuberculosis detection directly. To overcome these problems, we adopt two strategies. We present overlapping sub-image partition strategy for the small-size problem caused by anchor structure which is component of prevalent CNN-based object-detection method. The partition strategy overlappingly partition raw image into sub-images as per a formula presented by us. After partitioning, the proportion of M. tuberculosis on input image of model have been increased, that improving detecting accurate but reducing detecting speed. According to practice, we deem it acceptable. By cooperating with the co-author, 13 261 labeled data of M. tuberculosis have been constructed. Through a series of experiments, it has proved that our method is effective not only in improving detecting accurate and generalization of the model, but also in reducing necessary labeled data. The methods have been integrated into medical inspection products and confirmed to satisfy practical application requirements.

Key words: small target detection, medical image, M. Tuberculosis, convolutional neural networks